Intelligent monitoring method for crop growth cycle based on multi-source data fusion

Abstract

This invention relates to the field of smart agriculture technology and discloses an intelligent monitoring method for crop growth cycles based on multi-source data fusion. The method includes acquiring infrared reflectance, microwave scattering coefficient, leaf photosynthetic yield, canopy reflectance index, and time-series accumulated temperature of plants in a given area; outputting ear enlargement and tissue density based on joint addressing constrained by physical morphology and physiological state; inputting multi-source parameters into a competitive model to obtain sugar content curves, turgor pressure variation rate, and aroma rate; constructing a signed distance field to extract structural features of the area, and combining multi-scale time-frequency domain feature decomposition to extract global memory features; performing look-ahead offset compensation alignment through temporal phase difference calculation to output plant development stage values; extracting topological evolution features of aroma rate to calibrate extreme points, determining physiological maturity status, and finally generating a harvest prescription map. This invention approximately decouples and aligns the asynchronous internal and external development processes of crops, achieving objective determination of the true development stage of crops.

Description

Technical Field

[0001] This invention relates to the field of smart agriculture technology, specifically to a method for intelligent monitoring of crop growth cycles based on multi-source data fusion. Background Technology

[0002] In modern agricultural management, accurate monitoring of crop growth cycles is the core foundation for achieving scientific water and fertilizer management, yield prediction, and timely harvesting. With the development of agricultural sensing technology, various monitoring methods are now widely used in the industry to obtain crop growth information. For example, visual images or radar technology are used to observe changes in the external physical structure and morphology of crops (such as plant height and fruit volume); simultaneously, accumulated temperature data is obtained through weather stations, or vegetation indices are acquired using spectroscopic instruments to assess crop growth trends and physiological environmental conditions. Through this multi-dimensional environmental and phenotypic data, agricultural managers can track the growth and development process of crops to a certain extent.

[0003] However, the growth and development of crops (especially the expansion and ripening process of fruit crops) is a highly complex dynamic biological system. Crop ripening is not only manifested in the expansion of external physical morphology (such as increased volume and denser tissue), but also depends on the evolution of internal physiological metabolism (such as sugar accumulation, changes in turgor pressure, and the synthesis of aroma compounds). In the complex environment of real farmland, influenced by factors such as light, temperature fluctuations, and local microenvironmental stress, the growth rhythm of the crop's external physical morphology and the rate of evolution of its internal physiological metabolism often exhibit nonlinear and asynchronous characteristics.

[0004] Current technologies for addressing this type of growth monitoring need typically analyze morphological monitoring data and physiological environmental indicators independently, or employ simple linear weighting models for coarse-grained assessments. This approach struggles to deeply characterize the nonlinear coupling relationship between physical morphology and internal metabolism in the spatiotemporal dimensions, and lacks an alignment mechanism capable of objectively quantifying and eliminating this "asynchrony between internal and external development."

[0005] In summary, the field currently faces a pressing technical problem: how to approximately decouple the asynchronous evolution of the external physical morphology and internal physiological metabolism of crops in a complex and dynamic farmland environment, so as to determine the true developmental stage and physiological maturity of crops.

[0006] To address this, a smart monitoring method for crop growth cycles based on multi-source data fusion is proposed. Summary of the Invention

[0007] This invention relates to the field of smart agriculture technology and discloses an intelligent monitoring method for crop growth cycles based on multi-source data fusion. The method includes acquiring infrared reflectance, microwave scattering coefficient, leaf photosynthetic yield, canopy reflectance index, and time-series accumulated temperature of plants in a given area; outputting ear enlargement and tissue density based on joint addressing constrained by physical morphology and physiological state; inputting multi-source parameters into a competitive model to obtain sugar content curves, turgor pressure change rate, and aroma rate; constructing a signed distance field to extract structural features of the area, and combining multi-scale time-frequency domain feature decomposition to extract global memory features; performing look-ahead offset compensation alignment through temporal phase difference calculation to output plant development stage values; extracting topological evolution features of aroma rate to calibrate extreme points, determining physiological maturity status, and finally generating a harvest prescription map. This invention achieves objective determination of the true development stage of the crop by decoupling and aligning the asynchronous internal and external development processes of the crop.

[0008] To achieve the above objectives, the present invention provides the following technical solution: Intelligent monitoring methods for crop growth cycles based on multi-source data fusion include: The infrared reflectance, microwave scattering coefficient, leaf photosynthetic yield, canopy reflectance index, and time-series accumulated temperature of the plants in the area were obtained. Based on the physical morphological constraints of microwave scattering coefficient and the physiological state constraints of infrared reflectivity, the ear enlargement and tissue density are output through joint boundary intersection addressing.

[0009] By inputting leaf photosynthetic yield, time-series accumulated temperature, and canopy reflectance index into a multi-parameter competitive model, the sugar content curve, turgor pressure change rate, and aroma rate of the fruit were obtained.

[0010] Based on the enlargement and density of the fruit cluster, a signed distance field is constructed to extract the structural features of the plant area. Based on the sugar content curve, turgor pressure change rate, and aroma rate of the plant area, global memory features are extracted through multi-scale time-frequency domain feature decomposition and multi-level feature fusion. Based on the structural features of the plant area and the global memory features, the developmental stage of the plant is calculated by time-series phase difference. Based on the aroma rate, a continuous time-series curve is constructed to extract the topological evolution features of the connected components of the aroma rate of the plant area, and the global maximum point of the aroma rate synthesis rate is calibrated to determine the physiological maturity status of the fruit.

[0011] Based on the developmental stage values ​​and the physiological maturity of the fruit, a harvest prescription map is generated to delineate the distribution of harvestable areas.

[0012] Preferably, the steps for obtaining the infrared reflectance, microwave scattering coefficient, leaf photosynthetic yield, time-series accumulated temperature, and canopy reflectance index of the plants in the area include: using an image sensor to obtain the infrared reflectance of the plants in the area; using a polarimetric synthetic aperture radar to release microwave pulse signals for canopy penetration detection to obtain the microwave scattering coefficient of the plants in the area; using an ultra-hyperspectral imager to perform continuous spectral scanning detection of the canopy of the plants in the area to obtain the leaf photosynthetic yield; using a multispectral camera to take overhead images to obtain the canopy reflectance index; and using an off-ground meteorological sensing terminal to collect air temperature data at a continuous sampling frequency and obtain the time-series accumulated temperature of the plants in the area through integral calculation based on a reference temperature.

[0013] Preferably, the process for outputting the ear enlargement and tissue density of plants in the region includes: extracting the structural scattering features of the microwave scattering coefficient, calling a preset crop spatial topology prior knowledge base to perform feature space mapping calculation on the structural scattering features, and delineating the physical morphological boundary set of plant development; extracting the spectral absorption features of the infrared reflectance, calling a preset prior knowledge base to perform feature space mapping calculation on the spectral absorption features, and delineating the physiological development boundary set of plant development; performing logical intersection calculation on the physical morphological boundary set and the physiological development boundary set in a unified feature phase space to obtain the effective development domain; parsing the steady-state node attributes contained in the effective development domain, and directly mapping and outputting the spatial dimension parameter and density parameter corresponding to the attribute as the ear enlargement and tissue density of the plants in the region.

[0014] Preferably, the multi-parameter competitive model includes: Metabolic potential encoding layer: Based on a preset historical time sliding window, extract the photosynthetic yield of the leaves and the accumulated temperature of the time series for the consecutive N days before the current time node to construct a one-dimensional time vector, and perform tensor splicing operation to construct the basic metabolic potential vector in the feature space; Hyperplane generation layer: Extract the scalar features of the canopy reflectance index, use the scalar features as spatial bias parameters and scale scaling factors, dynamically adjust the mapping scale of the environmental transformation matrix, generate the hyperplane corresponding to the transformation matrix in the feature space, and define the orthogonal projection boundary of the basal metabolic potential energy through the geometric position of the hyperplane. Orthogonal projection decoupling layer: The basic metabolic potential energy vector is spatially geometrically aligned with the hyperplane, and an orthogonal decomposition calculation oriented towards the hyperplane is performed on the basic metabolic potential energy vector to separate the parallel component vector and the vertical residual vector. Competitive allocation mapping layer: receives parallel component vectors and vertical residual vectors; inputs the parallel component vectors into a preset temporal prediction network and a first fully connected layer, and outputs the sugar content curve and turgor pressure variation rate; inputs the vertical residual vectors into a preset second fully connected layer, and outputs the aroma production rate.

[0015] Preferably, the step of extracting the structural features of the region includes: receiving the ear enlargement and tissue density of the plants in the region; defining the geometric bounding box boundary in the three-dimensional coordinate system by the ear enlargement, and performing grid sampling on the space within the boundary to generate discrete coordinate points; assigning the tissue density as a quality weight to the discrete coordinate points to construct a set of discrete sampling anchor points; based on the spatial distribution topology of the discrete sampling anchor points, continuously fitting the coordinate space using an implicit neural function to construct a signed distance field; performing differential calculation on the signed distance field to solve for the continuous gradient distribution of the distance field, and calculating the Gaussian curvature tensor based on the Jacobian matrix of the gradient distribution; performing feature dimension concatenation operation on the continuous gradient distribution and the Gaussian curvature tensor, and performing dimensionality reduction encoding to map and output the structural features of the region.

[0016] Preferably, the steps for obtaining the global memory features include: resampling and aligning the sugar content curves, turgor pressure change rate, and aroma rate of plants in the same area on the same time axis, and performing tensor splicing to obtain a metabolic time series signal; setting a multi-scale time sliding window for different phenological stages of the covered crop, and performing mean and variance operations on the metabolic time series signal to obtain the trend mean term and fluctuation variance term of physiological fluctuations; and using a discrete wavelet transform operator to perform multi-level decomposition and downsampling processing on the metabolic time series signal to separate the approximate coefficients corresponding to the growth trend and the detail coefficients corresponding to the environmental fluctuations. Discrete integral operations are performed on the feature channels corresponding to the sugar content curve and aroma rate in the approximation coefficients to obtain the long-term cumulative values ​​of sugar content and aroma substances; the feature channels corresponding to the turgor pressure variation rate in the detail coefficients are extracted and their amplitude sum of squares is calculated to obtain the short-term turgor pressure fluctuation energy value; the trend mean term, fluctuation variance term, the cumulative values ​​of sugar content and aroma substances and the short-term turgor pressure fluctuation energy value are vectorized and input into a preset principal component analysis model and linear mapping matrix for subspace projection, and the global memory feature is output after removing redundant dimensions.

[0017] Preferably, the step of outputting the developmental stage value of the current plant includes: inputting the regional structural features and the global memory features from the continuous time series into a preset time sliding window, performing partial derivative calculations on the corresponding multidimensional feature vectors and feature latent states, and outputting the morphological change gradient and metabolic evolution rate respectively; constructing a morphological-physiological asynchronous delay metric based on the developmental phase difference between the morphological change gradient and the metabolic evolution rate; using the morphological-physiological asynchronous delay metric to perform look-ahead offset compensation on the time axis of the regional structural features, so that the physical morphological features and physiological metabolic features are aligned on the same developmental phase plane; performing multidimensional topological mapping on the aligned regional structural features and the global memory features to extract co-evolutionary parameters; measuring the distance between the co-evolutionary parameters and the preset standard crop variety growth and development trajectory feature space, and outputting the developmental stage value of the plant in the region based on the measurement deviation.

[0018] Preferably, the step of determining the physiological maturity state includes: constructing a continuous time-series curve based on the aroma rate of the area; extracting the time series of the aroma rate of the area by dynamically adjusting the filtering threshold along the signal amplitude axis; generating a persistent barcode by calculating the connected component evolution characteristics of the sequence at different scales; retrieving the core feature item with the longest lifespan in the persistent barcode; marking the topological extinction time corresponding to the core feature item as the global maximum point of the synthesis rate; and determining the physiological maturity state of the fruit based on the time position of the maximum point.

[0019] Preferably, the steps of generating the harvest prescription map and delineating the distribution of harvestable areas include: mapping the acquired developmental stage values ​​and fruit physiological maturity status to a preset farm geographic information coordinate matrix to obtain a set of spatial nodes with physiological status labels; retrieving preset logistics constraints and fruit grading standards, performing joint screening and priority scoring calculation on the set of spatial nodes to obtain a harvest distribution map; extracting nodes with priority values ​​higher than a preset threshold from the harvest distribution map, and performing spatial neighborhood scanning on these nodes to statistically analyze local spatial cluster density, obtaining dense communities representing areas with consistent maturity and concentrated harvesting value; performing gridded coordinate calibration on the dense communities, and assigning harvest time labels according to priority order, outputting the harvest prescription map containing gridded spatial coordinates and batch harvesting priorities. Compared with the prior art, the beneficial effects of the present invention are: 1. In actual farmland environments, the external morphological expansion and internal physiological evolution of crops are easily disrupted by environmental stress, leading to rhythmic misalignment. By calculating the developmental phase difference between the morphological change gradient and the metabolic evolution rate, an asynchronous delay metric is introduced to perform look-ahead offset compensation on the time axis. This mechanism can bring physical and physiological characteristics back to the same baseline at the dynamic evolution level, effectively correcting the asynchronous error in internal and external development caused by environmental disturbances, making the final output crop development stage values ​​more consistent with the actual biological state.

[0020] 2. To address the complex resource allocation mechanisms within crops, this invention constructs a unique competitive model. A hyperplane is generated using the canopy reflectance index as a bias parameter. The "basal metabolic potential vector," composed of photosynthetic yield and time-series accumulated temperature, is orthogonally decomposed toward the hyperplane, separating the parallel component (dominantly affecting turgor pressure) and the perpendicularly orthogonal component (dominantly affecting aroma rate). This method approximately decouples the intrinsic competitive relationships in nutrient synthesis during crop growth, enabling the prediction of sugar content, turgor pressure, and biochemical rates within the fruit.

[0021] 3. In the maturity determination stage, a "persistent barcode" representing the fluctuation trajectory of aroma synthesis rate is generated, and the core feature with the longest life cycle is retrieved to mark the global maximum point. This method objectively anchors the "optimal physiological maturity point" (i.e., optimal flavor period) of the fruit from the perspective of the global dynamic evolution of molecular synthesis rate; at the same time, combined with farm geographic information, logistical constraints, and fruit grading standards, the algorithm results are directly transformed into a "harvest prescription map" that includes batch harvesting priorities, providing guidance for large-scale agriculture. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the intelligent monitoring method for crop growth cycle based on multi-source data fusion according to the present invention. Figure 2 This is a flowchart illustrating the multi-parameter competitive model of the present invention; Figure 3 This is a schematic diagram illustrating the process of generating a harvest prescription map and delineating the distribution of harvestable areas in this invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Please see Figures 1 to 3This invention provides an intelligent monitoring method for crop growth cycle based on multi-source data fusion, the technical solution of which is as follows: A smart monitoring method for crop growth cycle based on multi-source data fusion, specifically as follows: Figure 1 As shown, it includes: The infrared reflectance, microwave scattering coefficient, leaf photosynthetic yield, canopy reflectance index, and time-series accumulated temperature of the plants in the area were obtained. Based on the physical morphological constraints characterized by the microwave scattering coefficient and the physiological state constraints of the infrared reflectivity, the ear enlargement and tissue density are output through joint boundary intersection addressing.

[0025] Based on the leaf photosynthetic yield, the accumulated temperature over time, and the canopy reflectance index, a multi-parameter competitive model is used to obtain the sugar content curve, turgor pressure change rate, and aroma rate of the fruit at the current time point.

[0026] Based on the enlargement and density of the fruit cluster, a signed distance field is constructed to calculate the Gaussian curvature tensor to extract the structural features of the region. Based on the sugar content curve, turgor pressure variation rate, and aroma rate of the region, global memory features are extracted through multi-scale time-frequency domain feature decomposition and multi-level feature fusion. Based on the calculation of temporal phase difference, benchmark alignment and co-evolutionary mapping are performed on the regional structural features and global memory features to output the developmental stage value of the current plant. The topological evolution features of the connected components of the aroma rate time series of the region are extracted, the global maximum point of the synthesis rate is identified, and the physiological maturity status of the fruit is determined.

[0027] Based on the developmental stage values ​​and the physiological maturity status of the fruit, a harvest prescription map containing grid spatial coordinates and batch harvesting priorities is generated to delineate the distribution of harvestable areas.

[0028] Example 1: Further, the steps for obtaining the infrared reflectance, microwave scattering coefficient, leaf photosynthetic yield, time-series accumulated temperature, and canopy reflectance index of the plants in the area include: using an image sensor to obtain the infrared reflectance of the plants in the area; using a polarimetric synthetic aperture radar to release microwave pulse signals for canopy penetration detection to obtain the microwave scattering coefficient of the plants in the area; using an ultra-hyperspectral imager to perform continuous spectral scanning detection of the canopy of the plants in the area to obtain the leaf photosynthetic yield; using a multispectral camera to take overhead images to obtain the canopy reflectance index; and using an off-ground meteorological sensing terminal to collect air temperature data at a continuous sampling frequency and obtain the time-series accumulated temperature of the plants in the area through integral calculation based on a reference temperature.

[0029] In the process of obtaining infrared reflectance, an indium gallium arsenide (IGaAs) image sensor with a short-wave infrared response band is selected and installed on the top bracket of a mobile inspection platform. The mobile inspection platform cruises through the field furrows according to a preset daily sampling cycle. During the journey, the IGaAs image sensor performs continuous spectral scanning and detection of the plant canopy below. The acquired raw spectral image data is transmitted to the vehicle-mounted control terminal. After the terminal performs dark current subtraction and distortion frame removal on the image, it outputs the absolute infrared reflectance data and records the corresponding spatial timestamp.

[0030] In the process of obtaining microwave scattering coefficients, a polarimetric synthetic aperture radar system is mounted on the gimbal of an unmanned aerial vehicle (UAV). Before the operation, a gridded mapping route covering the target area is planned. When the UAV flies along the set route, the radar releases microwave pulse signals to the canopy of the plants below and receives multipath scattering echo signals from the internal tissues of the canopy and the ground surface, thereby extracting the microwave scattering coefficients of the plants in the area.

[0031] In the process of obtaining leaf photosynthetic yield, an ultra-hyperspectral imager mounted on an unmanned aerial vehicle platform is used to continuously scan and detect the canopy of plants in the area. The equipment extracts the Fraunhofer dark line band in the reflectance spectrum, and by analyzing the canopy absorption depth filling effect, outputs the intensity of sunlight-induced chlorophyll fluorescence radiation excited by actual photosynthesis. Simultaneously, environmental incident photosynthetically active radiation data collected by an off-ground meteorological sensing terminal are retrieved, and the ratio of the sunlight-induced chlorophyll fluorescence radiation intensity to the environmental incident radiation is calculated and scaled to output the current leaf photosynthetic yield of the area.

[0032] In the process of obtaining the canopy reflectance index, a multispectral camera is mounted on an unmanned aerial vehicle (UAV). The operation time is limited to around noon. The UAV takes images at a fixed altitude along the flight path. After the flight, the processing system performs feature point extraction and matching, aerial triangulation, and radiometric correction on the acquired image sequence in sequence, and stitches them together to generate a multispectral orthophoto image covering the target area. Then, the reflectance pixel values ​​of the corresponding target feature bands in the orthophoto image are extracted, and algebraic operations based on the normalized ratio of band difference and band sum are performed to generate the continuous canopy reflectance index of the area.

[0033] In the implementation process of obtaining the time-series accumulated temperature of plants in a monitoring area, an off-ground meteorological sensing terminal is installed at the geographical center of the monitoring area. The data acquisition module of the sensing terminal reads the surrounding air temperature analog signal in real time, and after analog-to-digital conversion, the digital temperature value is recorded locally. The processor calculates the arithmetic mean of all the discrete temperature data collected continuously within a single day to obtain the daily average temperature. Then, the daily average temperature is compared with the preset biological lower limit benchmark temperature for crop varieties. If the daily average temperature is higher than the benchmark temperature, the temperature difference between the two is calculated as the effective temperature for that day. Finally, along the time series, all the daily effective temperature values ​​within the continuous growth cycle are accumulated and integrated day by day to output the time-series accumulated temperature of the area at the current node.

[0034] This multi-source data sensing system constructs a three-dimensional agricultural monitoring system. By using unmanned aerial vehicles, ground mobile inspection platforms, and fixed meteorological terminals equipped with various physical sensors, it collects five major environmental and physiological indicators of plants: infrared reflectance, microwave scattering coefficient, leaf photosynthetic yield, canopy reflectance index, and time-series accumulated temperature. By integrating multi-frequency multi-source heterogeneous sensing technologies and using automated data analysis algorithms to bind spatiotemporal correlation information, it achieves the monitoring of crop growth status and environmental changes.

[0035] Furthermore, the process for outputting the ear enlargement and tissue density of plants in the region includes: extracting the structural scattering features of the microwave scattering coefficient, calling a preset crop spatial topology prior knowledge base to perform feature space mapping calculation on the structural scattering features, and delineating the physical morphological boundary set of plant development; extracting the spectral absorption features of the infrared reflectance, calling a preset prior knowledge base to perform feature space mapping calculation on the spectral absorption features, and delineating the physiological development boundary set of plant development; performing logical intersection calculation on the physical morphological boundary set and the physiological development boundary set in a unified feature phase space to obtain the effective development domain; parsing the steady-state node attributes contained in the effective development domain, and directly mapping and outputting the spatial dimension parameter and density parameter corresponding to the attribute as the ear enlargement and tissue density of the plants in the region.

[0036] In the implementation process of extracting the physical morphological boundary set, the microwave scattering coefficient is decoupled into surface scattering, dihedral scattering, and volume scattering components using a "three-component decomposition algorithm" based on the physical scattering mechanism. The amplitudes of these three components are then concatenated dimensionally to construct a structural scattering feature vector. Subsequently, a pre-defined crop spatial topology prior knowledge base is established. This knowledge base, with a Gaussian mixture model as its core, pre-stores the standard spatial structure probability distribution parameters of this type of crop variety at different developmental stages. The specific process of establishing this prior knowledge base is as follows: During the complete historical growth cycle of the crop, three-dimensional point cloud data and multi-polarized microwave data of the plant are collected simultaneously; the three-dimensional point cloud is divided into voxel grids, and the density of the actual plant tissue point cloud within each voxel is statistically analyzed as the actual spatial occupancy probability; using the microwave scattering feature vector of the same developmental stage as a condition variable, and the corresponding actual spatial structure probability distribution parameters are then analyzed. The occupancy probability is set as the target distribution. A Gaussian mixture model is trained using the expectation-maximization algorithm, and the mean, covariance matrix, and mixture weight parameters of the Gaussian distribution are iteratively optimized. When the structural scattering feature vector is input into the model, the model outputs the three-dimensional Gaussian mixture distribution parameters (including the mean, covariance matrix, and weights of each Gaussian component) for the plant based on the learned parameters. Then, a Gaussian mixture probability density function with the three-dimensional coordinate system as the independent variable is constructed to calculate the occupancy probability distribution of each part of the plant on the three-dimensional spatial grid points. The structural scattering feature vector is input into the model to calculate the occupancy probability distribution of each part of the plant in the three-dimensional space. Finally, a structural confidence threshold is set (95% in this example), and the set of coordinates of three-dimensional spatial grid points with probability values ​​greater than the threshold is extracted to form a physical morphological boundary set that envelops the physical structural contour of the current plant.

[0037] In the implementation process of extracting the physiological development boundary set, the reflectance spectral sequence is first processed using the continuum removal method to extract the core absorption valley depth and absorption area, which characterize the water and dry matter accumulation in the plant, thus forming a spectral absorption feature vector. A pre-set prior knowledge base is then invoked, which contains a radial basis function neural network model trained based on a large number of plant physicochemical indicators and spectral features. The training method for this radial basis function neural network model is as follows: real water content and dry matter percentage of different canopy heights and different organ tissues (such as leaves, stems, and fruit spikes) are obtained through laboratory sampling beforehand, serving as physicochemical label data. The synchronously acquired spectral absorption feature vector is used as the input layer parameter, and the aforementioned physicochemical label data is stratified and labeled according to plant canopy height, serving as the output layer parameter. Specifically, the radial basis function neural network uses a Gaussian radial basis kernel function, with a kernel width parameter... In this example, the goal is to determine the target by performing 10-fold cross-validation on the training set and minimizing the moisture content prediction error. The value range is 0.1 to 0.5. To ensure the model's generalization ability, the training sample collection scale is set to no less than 200 plants, and the sampling needs to cover typical sample data of the target crop variety at different phenological stages such as seedling stage, swelling stage, and maturity stage. The network is trained and optimized using mean squared error as the loss function, so that the network can map the nonlinear relationship between spectral features and biochemical indicators at different canopy heights. After inputting the current spectral absorption feature vector into the trained neural network model, the model outputs the water content and dry matter accumulation values ​​at different heights of the upper, middle, and lower layers of the plant, thus forming biochemical concentration distribution data in the vertical spatial dimension. Specifically, by retrieving the standard biochemical parameter mode matching the current time series accumulated temperature stage from the preset prior knowledge base, the average water content of the crop variety at the corresponding development node is obtained. Mean of dry matter accumulation and its corresponding standard deviation , The spectral absorption feature vector is input into a neural network model to calculate the biochemical concentration distribution data such as water content and dry matter accumulation in the current tissue. Using the mean value as the center, a dynamic threshold window is constructed using a preset confidence increment coefficient k to calculate the physiologically reasonable range for each tissue level. The calculated biochemical concentration values ​​at each height position are compared with the physiologically reasonable range. If the value falls within the range, the coordinates of the position and its biochemical weight are retained. If it exceeds the range, the nearest neighbor interpolation method is used to correct the abnormal point to the range boundary value. The Alpha-shape algorithm is used to extract the minimum convex hull surface containing all points that meet the physiological threshold, and the physiological development boundary set of the plant is generated.

[0038] In the implementation process of obtaining the effective developmental domain, a dimensionless unified feature phase space was constructed. This unified feature phase space is a multi-dimensional attribute set that integrates morphological occupancy probability and physiological achievement probability, based on three-dimensional voxel coordinates. Specifically, the process involves: using a max-min normalization algorithm, aligning the spatial coordinate system of the physical morphological boundary set with the biochemical concentration range of the physiological developmental boundary set, mapping them to dimensionless probability distribution matrices within the same numerical range; and performing a spatial broadcast diffusion operation on the physiological developmental boundary set. Specifically, using the biochemical concentration distribution data in the vertical spatial dimension as a reference, Gaussian smoothing interpolation is performed on the concentration values ​​in the horizontal direction to expand and fill the range to match the physical morphological boundary set. In a three-dimensional voxel grid with completely consistent morphological boundary sets, each voxel grid in the spatial probability distribution matrix stores the probability value of the morphology, while the same voxel grid coordinate in the biochemical probability distribution matrix stores the normalized probability value of the biochemical concentration at that location that conforms to a reasonable physiological state, making the two matrices have completely consistent three-dimensional row, column, and page dimensional structures. In a unified phase space, element-wise multiplication operations are performed on the spatial probability distribution matrix and the biochemical probability distribution matrix to achieve logical intersection. A connected component analysis algorithm is used to extract closed high-dimensional regions in the product matrix whose values ​​are continuous and higher than a preset extreme value threshold. The closed region after multiple attribute overlap verification is marked as the current effective developmental domain of the plant.

[0039] The preset extreme value threshold is obtained by dynamically adjusting the initial threshold generated by the Otsu algorithm and the fault tolerance factor set based on the current phenological period through the statistical histogram of the product matrix, as shown in the following formula: ; in, The preset extreme value threshold; To use the Otsu algorithm to search for the critical value that maximizes the inter-class variance of the histogram, as the initial segmentation threshold; To determine the tolerance coefficient based on the historical developmental dispersion of the target crop variety during the current phenological period, specifically, the standard deviation of the effective developmental domain product matrix of historical samples within this phenological period is calculated. ,like If the percentage is greater than 15% of the mean, then γ = 0.2; if it is less than 5%, then γ = −0.1; otherwise, linear interpolation is performed between the two values. When the calculated result exceeds [0, 1], it is truncated to the boundary value. The final value is limited to the range [0, 1].

[0040] In the process of outputting ear enlargement and tissue density, the mean-shift clustering algorithm is used to find data point clusters with zero probability density gradient and the highest local joint probability density within the effective developmental domain. These local extreme centers are marked as steady-state nodes. Then, the volume span of the maximum bounding box of the core region of the steady-state node on the physical mapping axis is used as a spatial dimension parameter; the historical average bounding box volume of the corresponding crop variety at the initial pollination period in the database is used as the baseline volume. Subsequently, the currently calculated spatial dimension parameter is divided by this baseline volume to calculate the volume change factor, which is directly output as the ear enlargement value. Simultaneously, the biochemical concentration-weighted average of the steady-state node on the physiological mapping axis is used as a density parameter, and a preset variety calibration conversion coefficient table is retrieved to linearly map it to the final tissue density value of the target part for output.

[0041] By using spatial broadcast diffusion operations and constructing a unified dimensionless feature phase space, the computational fusion of heterogeneous data between three-dimensional physical structure features and one-dimensional biochemical concentration features was completed, realizing the alignment of multimodal sensing parameters and verification of multiple attributes. By quantitatively extracting "ear enlargement" and "tissue density" that truly reflect the actual growth and development state of crops, data support is provided for the subsequent generation of agricultural harvest prescription maps.

[0042] Furthermore, the multi-parameter competitive model includes, specifically, as follows: Figure 2 As shown: Metabolic potential encoding layer: Tensor splicing operation is performed on the photosynthetic yield of the leaf and the accumulated temperature of the time series to construct the basic metabolic potential vector in the feature space; Hyperplane generation layer: Extract the scalar features of the canopy reflectance index, use the scalar features as spatial bias parameters and scale scaling factors, dynamically adjust the mapping scale of the environmental transformation matrix, generate the hyperplane corresponding to the transformation matrix in the feature space, and define the orthogonal projection boundary of the basal metabolic potential energy through the geometric position of the hyperplane. Orthogonal projection decoupling layer: Spatially align the basal metabolic potential energy vector with the hyperplane, perform orthogonal decomposition calculations toward the hyperplane on the basal metabolic potential energy vector, and separate the parallel component vector parallel to the hyperplane and the vertical residual vector perpendicular to the hyperplane. Competitive allocation mapping layer: receives parallel component vectors and vertical residual vectors; inputs the parallel component vectors into a preset temporal prediction network and a first fully connected layer, and outputs the sugar content curve and turgor pressure variation rate; inputs the vertical residual vectors into a preset second fully connected layer, and outputs the aroma production rate.

[0043] In the implementation process of the metabolic potential coding layer, the processing unit first performs temporal augmentation on the acquired leaf photosynthetic yield, extracting historical photosynthetic yield data for the consecutive N days prior to the current time point to form a one-dimensional time-series vector; simultaneously, it extracts the corresponding time-series accumulated temperature vector for the corresponding consecutive N days. Tensor concatenation is then performed on the feature channels of these two vectors with the same time length dimension to construct a two-dimensional matrix sequence M. The determination principle for the value of the historical time sliding window length N includes: 1. Dynamic setting method: The basal metabolic potential vector is dynamically set according to the average duration of the current phenological stage (such as flowering stage, fruit enlargement stage) of the target crop variety, so that the basal metabolic potential vector can fully cover the metabolic evolution trend of this physiological stage. 2. Fixed Empirical Method: A fixed value is set based on the experience of agricultural experts, preferably 15 to 45 days (e.g., 30 days), to balance computational complexity and the richness of feature representation. To eliminate the differences in physical dimensions and numerical scales between leaf photosynthetic yield and time-series accumulated temperature, before performing tensor splicing, the two one-dimensional time-series vectors are respectively subjected to z-score standardization based on statistical parameters of the historical training set. Then, flattening and splicing operations are performed to construct the basal metabolic potential vector. .

[0044] In the implementation process of the hyperplane generation layer, the hyperplane generation layer system has a pre-set basic environment transformation matrix obtained by offline training using historical environment datasets. The specific offline training process for establishing this basic environmental transformation matrix is ​​as follows: An offline dataset containing multidimensional environmental parameters (such as temperature, humidity, and light radiation) under different phenological periods over the years is constructed; principal component analysis is used to orthogonally transform the feature space of the environmental parameters in the offline dataset, extracting the leading edge parameters that can characterize the direction of maximum change in environmental variance. Each principal component eigenvector; [This] The principal component feature vectors are combined column-wise to construct a basic environment transformation matrix covering the basic environment mapping dimension. After the processing unit extracts the canopy reflectance index, it performs a global average pooling operation to obtain a scalar feature value c. The system inputs this scalar feature value into a preset parameter generation subnetwork, which contains a shared fully connected hidden layer and two parallel independent output branches; the first output branch uses the Sigmoid activation function to output a k-dimensional positive vector. (k is the number of principal components). Each component corresponds to a basic environment transformation matrix. One column direction enables independent scaling of each principal component direction; the second output branch uses the Tanh activation function to generate the bias matrix. (d is the dimension of the feature space), with Construct a diagonal matrix for the diagonal elements ,right Each column is scaled independently and then the offsets are superimposed to obtain the dynamic environment transformation matrix. : ; in, This is the dynamic environment transition matrix under the current environmental state; This is the basic environment transformation matrix obtained through offline training; For Construct a diagonal matrix for the diagonal elements. For the first branch output, This is the output for the second branch.

[0045] Due to the bias matrix The introduction of disrupted To ensure orthogonality, QR decomposition is required to extract the orthonormal basis U. Where Q is an orthogonal matrix and R is an upper triangular matrix. U satisfies Each column vector of the hyperplane forms an orthonormal basis for the hyperplane in the feature space. The geometric normal vector of the hyperplane defines the orthogonal projection boundary of the basal metabolic potential energy vector.

[0046] Since U already satisfies orthogonality, the orthogonal projection operator is simplified, and the basal metabolic potential vector is transformed. Decomposed into parallel component vectors With the vertical residual vector .

[0047]

[0048] ; in, It is the orthogonal projection of the basal metabolic potential vector onto the hyperplane, i.e., the parallel component vector; It is the vertical residual vector; It is the basal metabolic potential vector.

[0049] In the implementation process of the competitive allocation mapping layer, the competitive allocation mapping layer system receives the decoupled parallel component vectors. and vertical residual vector For the parallel component vector, it is synchronously input into the preset Long Short-Term Memory (LSTM) network (temporal prediction network) and the first fully connected layer respectively; the mapping hyperplane is dynamically generated through environmental features, and the total metabolic potential energy of the plant is orthogonally decomposed into two independent components representing basal growth and secondary metabolism; at the same time, the internal temporal gating calculation of the LSTM network is correlated with the real physiological processes of plant carbohydrate accumulation and environmental stress; the above orthogonal decomposition has a clear biological basis: the accumulation of soluble sugars and maintenance of turgor pressure during fruit development belong to primary metabolism, and its rate is highly positively correlated with the temporal accumulation of photosynthetic yield and the cell expansion driven by accumulated temperature; while terpenes, esters and other aroma substances belong to secondary metabolites, and their synthesis rate has a substrate competition relationship with primary metabolism; when the basal metabolic potential energy vector is decomposed to the hyperplane spanned by the environmental transformation matrix, the parallel component retains the direction collinear with the temporal change of photosynthetic yield and is consistent with the direction of the principal component of primary metabolic material flow; the perpendicular orthogonal component represents the remaining metabolic energy difference, corresponding to the metabolic flux outside of primary metabolism, that is, the intensity of secondary metabolism.

[0050] The parallel component vectors are input into a pre-defined Long Short-Term Memory (LSTM) network to extract temporal dependency features and output a sugar content curve. Before feature extraction, the structure and initial parameter configuration of the LTM network must be completed. Specifically, the feature dimension of the decoupled parallel component vectors is directly set as the node dimension of the LTM network input layer; simultaneously, a pre-defined hidden layer node dimension parameter is used to define the network cell state and hidden state vector h. t The parameter space capacity is determined. During parameter initialization, the weight matrices controlling the forget gate, input gate, and output gate operations within the network are assigned values ​​using an orthogonal initialization strategy. Furthermore, the initial value of the bias term corresponding to the forget gate is explicitly set to a constant 1, prompting the network to retain long-term basal metabolic history memory during the initial iteration phase, and the bias terms of the remaining gated units are initialized to 0. After completing the network structure establishment and parameter initialization, during the forward propagation process within this long short-term memory network: the forget gate structure is used to evaluate the hidden state of the previous time step and the parallel component input of the current time step, and the forgetting weight is calculated; the input gate structure is used to filter the effective basal nutrient accumulation trend in the current parallel component and update and accumulate it into the cell state of the network, so that the cell state can physically represent the long-term carbohydrate potential energy accumulation during the fruit development cycle; finally, the output gate structure outputs the hidden state vector of the current time step based on the updated cell state, which is a feature representation that highly condenses the physical growth time-dependent law, and the time-series prediction network directly outputs the sugar content curve.

[0051] The time series prediction network uses a single-layer LSTM with a hidden layer node dimension of 256 and an initial forget gate bias of 1. The first fully connected layer has 64 nodes, uses ReLU activation, and its output dimension matches the time length of the sweetness curve, with a time granularity of calendar days. The second fully connected layer has 32 nodes, uses Sigmoid activation, and outputs a scalar sweetness rate normalized to [0,1]. The joint loss function consists of three terms, all using the mean squared error function. , , The weighting coefficients for each loss are set in this embodiment after normalization based on the magnitude of each output value. , , =1:1:1 (or adjust according to the experiment).

[0052] The parallel component vectors are synchronously input into the first fully connected layer for nonlinear mapping, outputting the turgor pressure change rate, where the turgor pressure change rate represents the rate of change of turgor pressure relative to the reference value over time, corresponding to the first-order difference sequence of continuous sensor measurements as the training label. For the vertical residual vector, it is directly input into the preset second fully connected layer sequence decoder for nonlinear mapping, outputting the aroma production rate time sequence with the same time length as the training label. Before implementing the aforementioned network model, an end-to-end supervised training set needs to be constructed. To this end, leaf photosynthetic yield, time-series accumulated temperature, and canopy reflectance index under historical cycles are used as initial data. Following the same encoding and orthogonal projection decoupling rules, the parallel component vector and vertical residual vector under historical conditions are calculated as training input features for the network. Subsequently, real sugar content sequence data obtained by continuous measurement using a refractive index saccharimeter under historical cycles are used as sugar content labels. Real cell hydraulic pressure sensor continuous data measured at the same historical period of the plant are used as turgor pressure labels. For discrete aroma substance concentration data measured by low-frequency gas chromatography-mass spectrometry, Gaussian process regression algorithm is used to perform smooth interpolation along the time axis to generate a continuous time series with the same time resolution as the turgor pressure label as the aroma rate label. Using the aligned labels as supervision signals, the joint loss function L is minimized using the backpropagation algorithm to complete network supervision training and parameter solidification. In order to make the three losses comparable in magnitude, before calculating each MSE loss, the predicted values ​​and label values ​​of the sugar content curve and the rate of change of turgor pressure are normalized by z-score based on the training set statistics. The aroma rate has been normalized to [0, 1] and does not require additional processing.

[0053] By dynamically generating a mapping hyperplane based on environmental features, the total metabolic potential of the plant is orthogonally decomposed into two independent components representing basal growth and secondary metabolism. At the same time, the internal temporal gating calculation of the LSTM network is correlated with real physiological processes such as carbohydrate accumulation and environmental stress in plants, giving physical meaning to the feature extraction process. Finally, supervised training is performed using real historical data measured by physicochemical sensors to achieve quantitative assessment and interpretable prediction of fruit turgor pressure change rate, sugar content curve, and aroma rate.

[0054] Further, the step of extracting the structural features of the region includes: receiving the ear enlargement and tissue density of the plants in the region; defining the geometric bounding box boundary in the three-dimensional coordinate system by the ear enlargement, and performing grid sampling on the space within the boundary to generate discrete coordinate points; assigning the tissue density as a quality weight to the discrete coordinate points to construct a set of discrete sampling anchor points; based on the spatial distribution topology of the discrete sampling anchor points, continuously fitting the coordinate space using an implicit neural function to construct a signed distance field; performing differential calculation on the signed distance field to solve for the continuous gradient distribution of the distance field, and calculating the Gaussian curvature tensor based on the Jacobian matrix of the gradient distribution; performing feature dimension concatenation operation on the continuous gradient distribution and the Gaussian curvature tensor, and performing dimensionality reduction encoding to map and output the structural features of the region.

[0055] In the implementation process of receiving the ear enlargement and tissue density of plants in a receiving area and generating a discrete sampling anchor point set, a three-dimensional virtual Cartesian coordinate system is first established in memory. The received ear enlargement values ​​are extracted and used as the scale expansion reference. A physical spatial extension is then performed along the three orthogonal axes of the coordinate system, centered on the origin, defining a geometric bounding box that completely encloses the target ear volume. Subsequently, a fixed spatial grid resolution step size is set, and the three-dimensional space inside this geometric bounding box is uniformly meshed. The three-dimensional coordinates of all grid intersections are extracted as discrete coordinate points. Next, the received tissue density values ​​are used as a one-dimensional quality weight scalar and bound to the three-dimensional spatial coordinates of each discrete coordinate point in a data structure, forming a discrete sampling anchor point set containing four dimensions.

[0056] In the implementation process of constructing the signed distance field, a continuous spatial geometric field is reconstructed using a set of discrete sampling anchor points. Specifically, a multilayer perceptron network is pre-built as an implicit neural function. This implicit neural function adopts a 5-layer fully connected multilayer perceptron structure, and the training strategy of the multilayer perceptron network of the implicit neural function combines offline pre-training with online fast fine-tuning. The system pre-trains a general geometric prior network that captures the common geometric shape of crop varieties on a high-performance server using massive amounts of historically collected 3D point cloud data of fruit ears, and solidifies its model weights as the initial state for deployment to mobile terminals. The number of nodes in each layer is set to 256, and the activation function is a sine function to enhance the ability to express high-frequency geometric details on the fruit surface. Its input layer receives 3D coordinate data, and the output layer outputs a single distance scalar. A surface boundary threshold for tissue density is set based on prior physics. The specific methods for determining this threshold include: retrieving the historical average tissue density of the crop variety at different developmental stages from a preset prior knowledge base as a benchmark; or determining the critical density threshold at which the fruit epidermis and the fruit mesenchymal layer undergo abrupt changes through in vitro tissue mechanical property experiments, and setting it as the surface boundary threshold. Points with a density equal to the threshold in the discrete sampling anchor point set are marked as zero isosurfaces; points with a density greater than the threshold are considered inside the entity and assigned negative distance labels, while points with a density less than the threshold are considered outside the entity and assigned positive distance labels. All discrete coordinate points marked as zero isosurfaces are extracted, and a KD-tree spatial retrieval structure is constructed. For each discrete coordinate point inside and outside the entity, the KD-tree algorithm is used to search for the nearest zero isosurface coordinate point (the number of nearest neighbors K is set to 1), and the absolute value of the spatial Euclidean distance (i.e., L2 norm) between them is calculated. The absolute value of the Euclidean distance is multiplied by the aforementioned positive or negative distance labels to generate a true signed distance value with precise numerical value and sign. Using the aforementioned 3D coordinate points with distance labels as the supervised dataset, the multilayer perceptron network is iteratively trained to optimize the network weights. Specifically, the training loss function of the iterative training consists of three weighted components: a surface constraint term (i.e., the mean square error of the distance residual at SDF=0, with a weight of 1.0), a normal consistency constraint term (i.e., the cosine distance between the gradient vector and the true normal vector, with a weight of 0.1), and a spatial regularization term (i.e., the Eikonal constraint for non-surface points, with a weight of 0.1). After training convergence, the multilayer perceptron network can fit the shortest signed distance from any continuous coordinate point in the input space to the surface of the ear of fruit, thus forming a signed distance field.

[0057] In the implementation process of extracting the continuous gradient distribution and Gaussian curvature tensor, spatial geometric differentiation is performed on the aforementioned signed distance field; the first-order partial derivatives of the three-dimensional spatial coordinate variables input to the multilayer perceptron are calculated to obtain the set of normal vectors at each coordinate point in the three-dimensional space, forming a continuous gradient distribution; the second-order partial derivatives of the spatial coordinate variables are calculated again for the normal vector result of the continuous gradient distribution to solve for the Jacobian matrix of the gradient vector, which is equivalent to the Hessian matrix of the signed distance field; based on the extracted Hessian matrix, its eigenvalues ​​are calculated to obtain the principal curvatures, and the product of the principal curvatures is calculated to generate the Gaussian curvature tensor. It should be noted that the Gaussian curvature referred to here as the product of the two principal curvatures is a scalar field; at the same time, the average curvature is calculated, and the two are concatenated to form the local curvature descriptor of each coordinate point, which participates in the subsequent feature dimension concatenation operation; in the implementation process of extracting and outputting the regional structural features, the obtained continuous gradient distribution and the Gaussian curvature tensor at the corresponding coordinate position are concatenated point by point along the feature channel dimension to generate a high-dimensional joint geometric feature matrix. The joint geometric feature matrix is ​​input into a pre-defined dimensionality reduction encoder, which consists of offline pre-trained downsampling convolutional layers and pooling layers connected in series. The specific offline pre-training process of the dimensionality reduction encoder is as follows: a historical sample database for this type of crop variety is pre-constructed; the three-dimensional geometric feature matrix samples in the database are specifically derived from: collecting three-dimensional point cloud reconstruction models of this crop variety covering various typical phenological stages (such as seedling stage, jointing stage, heading stage, etc.) over multiple complete growth cycles in the past, and offline batch-generating three-dimensional grid spatial feature data containing continuous gradient distribution and Gaussian curvature tensor based on the same distance field construction and spatial partial derivative calculation rules; a three-dimensional autoencoder network containing the dimensionality reduction encoder and the corresponding reconstruction decoder is constructed; using the geometric feature matrix in the historical sample database as input, and aiming to reconstruct the original input features losslessly through the network output, the autoencoder is subjected to unsupervised iterative training using the mean square error loss function; after training convergence, the decoder part is discarded, and the network parameters of the dimensionality reduction encoder are solidified. The dimensionality reduction encoder performs local feature pooling and dimensionality compression operations on the high-dimensional joint geometric feature matrix, filters out redundant local spatial coordinate information, and maps it to a compact one-dimensional hidden state feature vector. This hidden state feature vector is then extracted and output as the current patch structure feature.

[0058] By utilizing a multilayer perceptron and KD-tree to transform discrete crop parameters into a continuously differentiable spatial distance field, the resolution limitations of traditional meshes are reduced. The Gaussian curvature tensor is resolved by an automatic differentiation engine of a deep learning framework, reducing the numerical approximation error of traditional discrete difference and quantifying the local geometric topology of the crop. Finally, dimensionality reduction is performed by an unsupervised pre-trained 3D autoencoder based on historical periodic point cloud data, achieving low-dimensional analysis of the crop's 3D structure.

[0059] Further, the steps for obtaining the global memory features include: resampling and aligning the sugar content curves, turgor pressure variability, and aroma production rate of plants in the same area on the same time axis, and performing tensor splicing to obtain a metabolic time series signal; setting a multi-scale time sliding window for different phenological stages of the covered crop, and performing mean and variance operations on the metabolic time series signal to obtain the trend mean and fluctuation variance terms of physiological fluctuations; using a discrete wavelet transform operator to perform multi-level decomposition and downsampling processing on the metabolic time series signal to separate the approximation coefficients corresponding to the growth trend and the detail coefficients corresponding to the environmental fluctuations; performing discrete integral operations on the approximation coefficients to obtain the long-term cumulative values ​​of solids and aroma substances; calculating the sum of squared amplitudes of the detail coefficients to obtain the short-term turgor pressure fluctuation energy value; concatenating the trend mean, fluctuation variance, cumulative values, and short-term turgor pressure fluctuation energy value into a vector, and inputting it into a preset principal component analysis model and linear mapping matrix for subspace projection, removing redundant dimensions, and outputting the global memory features.

[0060] In the process of acquiring metabolic time series signals, based on the sugar content curve, turgor pressure change rate, and aroma production rate data of different regions, the processing unit establishes a unified reference time axis with a set hour or natural day as the basic time resolution; the cubic spline interpolation algorithm is used to eliminate the data acquisition time difference between the sensors; the three aligned sequences are tensor-joined along the feature dimension to construct a two-dimensional matrix sequence, which is used as the metabolic time series signal.

[0061] In the implementation process of extracting short-term physiological fluctuations, based on the duration characteristics of the target crop variety at different phenological stages (such as jointing stage, flowering stage, and fruit enlargement stage), a multi-scale time sliding window with short, medium, and long spans is pre-set. The window span is strictly bound to the physiological rhythm of the plant: the time scale of the short span window is set to a 24-hour diurnal cycle to capture the diurnal rhythm fluctuations of plant photosynthetic metabolism caused by diurnal temperature differences and instantaneous sunlight; the time scale of the medium span window is set to a 7- to 15-day weekly cycle to capture the persistent fluctuations caused by local weather systems such as continuous rain or short-term drought. The continuous environmental stress fluctuations; the time scale of the long-span window is dynamically aligned with the complete duration of the current phenological period (if the fruit enlargement period) to capture macroscopic basic developmental trends such as the long-term transfer of carbohydrates to the fruit end; the processing unit translates and scans these sliding windows along the time axis of the metabolic time series signal matrix according to a set fixed time step; within the window interval extracted in each translation, the arithmetic mean and statistical variance of various metabolic characteristic data series are calculated using statistical algorithms; the extracted multi-scale mean sequence constitutes the trend mean term, and the extracted multi-scale variance sequence constitutes the fluctuation variance term.

[0062] In the implementation process of multi-scale time-frequency feature decomposition, the processing unit selects wavelet basis functions (in this embodiment, the db4 wavelet from the Daubechies wavelet family) to perform preset-level discrete wavelet decomposition processing on the above-mentioned metabolic time series signal. Since the pre-order tensor splicing operation preserves the original order of each channel (in this embodiment, feature channel indices 1, 2, and 3 correspond to the sugar content curve, turgor pressure change rate, and aroma rate, respectively), the discrete wavelet decomposition operation is performed independently for each feature channel along the time axis. In each level of decomposition, the signal sequence is simultaneously passed through a set of orthogonal low-pass and high-pass filters, and a half-sampling operation is performed on the filtered output data. After multi-level decomposition, the lowest frequency component of the final output of the low-pass filter branch is extracted as an approximation coefficient; while the high-frequency components output by each level of the high-pass filter branch are combined and extracted as detail coefficients.

[0063] In the implementation process of calculating the cumulative amount and mutation energy, the processing unit performs decoupled calculations on the multidimensional data channels based on the evolutionary differences in physiological characteristics: the processing unit extracts the feature channels corresponding to the sugar content curve and aroma rate from the above-separated approximation coefficients, and performs discrete numerical integration along the time axis using the trapezoidal integral rule, accumulating the low-frequency metabolic rate features in the discrete state over the time span, and calculates and outputs the long-term cumulative sugar content (i.e., solids) value and the cumulative aroma substance value; at the same time, the processing unit extracts the feature channels corresponding to the turgor pressure rate from the detail coefficients obtained from each layer decomposition, calculates the square value of each feature element in the sequence, and performs a summation operation in the corresponding time interval. The result of the square sum of the amplitude is physically equivalent to the time domain signal energy, and obtains the short-term turgor pressure fluctuation energy value, which characterizes the short-term fluctuation impact of environmental short-wave micro-perturbations (such as gusts, instantaneous cloud cover caused by micro-meteorological transients) on plant water metabolism.

[0064] In the implementation process of extracting and outputting global memory features, the processing unit sequentially performs a one-dimensional flattening operation on the aforementioned acquired trend mean, fluctuation variance, long-term cumulative value, and short-term fluctuation energy value of turgor pressure, and performs tensor concatenation along the feature dimension to construct a high-dimensional original memory vector covering multi-scale time-frequency information. The specific offline process for inputting this high-dimensional original memory vector into a pre-set principal component analysis model and establishing the principal component analysis model and covariance feature vector matrix is ​​as follows: Historical sugar content, turgor pressure, and aroma rate original sensor data of the target crop variety over its complete annual growth cycle are collected in advance; for the discrete sampling data obtained from the aroma rate original sensor, a Gaussian process regression algorithm is used to perform smooth interpolation processing along the time axis to align the time resolution of its output features with the historical sugar content and turgor pressure data; the historical original data undergoes the same baseline time axis resampling, multi-scale sliding window mean and variance calculation, and discrete wavelet processing as in the online monitoring stage. The process involves decomposition, discrete integration, and sum of squared amplitude operations to extract the historical trend mean, fluctuation variance, long-term cumulative values, and short-term turbulence energy values ​​corresponding to all samples over the years. These historical features are then flattened and concatenated in one dimension to construct an offline feature matrix composed of massive historical high-dimensional original memory vectors. The covariance matrix of this offline feature matrix is ​​calculated, and eigenvalue decomposition is performed on the covariance matrix to obtain the corresponding covariance eigenvector matrix. Simultaneously, the global feature mean vector of all historical high-dimensional original memory vectors is calculated and saved as parameters for the preset principal component analysis model. This model is based on the pre-calculated covariance eigenvector matrix of historical sample data. During projection, the currently acquired high-dimensional original memory vector is first subtracted from the pre-saved offline global feature mean vector to complete the mean-centered alignment of the feature data. Subsequently, the centered memory vector is multiplied linearly with the covariance eigenvector matrix to project it into an orthogonal low-dimensional feature subspace. Based on the preset cumulative variance contribution rate threshold (set to 95% in this example), the first K main feature columns after projection are retained, redundant feature dimensions with minimal variance contribution are removed, and finally, a compact vector after dimensionality reduction is output, which is used as the global memory feature of crops in the area.

[0065] This technical solution couples signal processing algorithms with crop physiological patterns. It reduces the acquisition time difference of heterogeneous data through interpolation and resampling to achieve multimodal parameter alignment, and uses a multi-scale sliding window that binds crop rhythms and phenological periods to extract time-domain physiological fluctuation features. With the help of discrete wavelet transform and calculus, the coupled signals are objectively decoupled into long-term material accumulation and short-term environmental energy in the frequency domain. Finally, based on the PCA model built on historical homogeneous data, online centering and orthogonal projection are performed to remove redundant space in high-dimensional features and obtain global memory features that reflect the dynamic evolution of crops.

[0066] Further, the step of outputting the developmental stage value of the current plant includes: inputting the regional structural features and the global memory features from the continuous time series into a preset time sliding window, performing partial derivative calculations on the corresponding multidimensional feature vectors and feature latent states, and outputting the morphological change gradient and metabolic evolution rate respectively; constructing a morphological-physiological asynchronous delay metric based on the developmental phase difference between the morphological change gradient and the metabolic evolution rate; using the morphological-physiological asynchronous delay metric to perform look-ahead offset compensation on the time axis of the regional structural features, so that the physical morphological features and physiological metabolic features are aligned on the same developmental phase plane; performing multidimensional topological mapping on the aligned regional structural features and the global memory features to extract co-evolutionary parameters; measuring the distance between the co-evolutionary parameters and the preset standard crop variety growth and development trajectory feature space, and outputting the developmental stage value of the regional plant based on the measurement deviation.

[0067] In the implementation process of calculating the morphological change gradient and metabolic evolution rate, based on the segmental structural feature vectors and global memory feature vectors stored continuously according to the timestamp sequence, the processing unit establishes a data buffer queue with a fixed time step as a preset time sliding window, so that the window slides synchronously along the time axis of the two feature sequences; for the continuous discrete feature vectors within the window, the central finite difference algorithm is used to calculate the first-order numerical partial derivatives of each dimension component of the structural feature vector and each dimension component of the global memory feature vector with respect to the time variable; the first-order partial derivatives of all structural feature components are recombined to output the morphological change gradient vector; at the same time, the first-order partial derivatives of all memory feature components are combined to output the metabolic evolution rate vector.

[0068] In the implementation process of constructing a morphological-physiological asynchronous delay metric, the processing unit extracts the morphological change gradient sequence and metabolic evolution rate sequence within the historical sliding window and performs developmental phase difference analysis in the time dimension. Specifically, using a cross-correlation analysis algorithm, with the metabolic evolution rate sequence as the main signal and the morphological change gradient sequence as the reference sliding signal, the cross-correlation coefficient sequence of the two under different discrete time lag steps is calculated. The cross-correlation coefficient sequence is traversed to find the global maximum value, and the lag time step value corresponding to the maximum value is extracted and quantified as a morphological-physiological asynchronous delay metric.

[0069] In the implementation process of look-ahead offset compensation and baseline alignment, based on the calculated asynchronous delay metric, the extracted regional structural features are subjected to time-axis data resampling and alignment operations. Specifically, the processing unit determines the time step value of the asynchronous delay metric: if it is an integer time step, it directly backtracks from the historical buffer queue and extracts the regional structural feature vector before that hysteresis step number; if it contains a decimal number of non-integer time steps, it extracts the regional structural feature vectors of two adjacent historical time nodes, and calculates the virtual structural feature vector at the non-integer hysteresis time using a linear interpolation algorithm; the historical regional structural feature vector after offset compensation is forcibly bound to the global memory feature vector at the current time with a timestamp, thereby eliminating the time difference of physiological conduction at the data level and enabling the physical morphological features and physiological metabolic features to achieve baseline alignment on the same developmental phase plane.

[0070] In the implementation process of extracting co-evolutionary parameters, the benchmark-aligned region structural feature vector and the global memory feature vector are subjected to tensor concatenation along the feature channel dimension to construct a joint state vector. Subsequently, this joint state vector is input into a pre-defined manifold learning topology mapping model. This mapping model, based on preset network weights, performs local spatial connectivity analysis and nonlinear dimensionality reduction on the input high-dimensional joint state vector, mapping it to a low-dimensional continuous manifold coordinate system. The specific offline training process for establishing and solidifying the manifold learning topology mapping model is as follows: A large number of historical benchmark-aligned joint state vectors of the target crop variety over its complete annual growth cycle are pre-collected to construct an unsupervised training set; a nonlinear autoencoder network containing an encoder and a decoder is constructed. The autoencoder consists of three fully connected layers (with 256, 128, and 32 nodes respectively) connected in series, using ReLU as the activation function, and the decoder has a symmetric structure; the low-dimensional manifold space dimension is set to 32; using the historical joint state vectors in the training set as input, and aiming to maximize the reconstruction of the input vector by the network output, iterative training is performed using the mean squared error loss function and backpropagation algorithm; after training convergence, the decoder is stripped, and the network weights of the encoder are extracted and solidified as preset network weights for online dimensionality reduction mapping. The mapped 32-dimensional low-dimensional coordinate vector is extracted and used as a co-evolutionary parameter characterizing the intensity of internal and external plant co-development.

[0071] In the implementation process of determining and outputting the current plant development stage value, a standard crop variety growth and development trajectory feature space is pre-constructed and stored in the database. This space contains a continuous standard trajectory curve generated by the same mapping model from standard superior crop samples over the entire life cycle. Each discrete anchor point on this curve is pre-associated with the actual crop development stage value (such as standard accumulated temperature days or maturity percentage). The specific construction process of the continuous standard trajectory curve is as follows: extract the low-dimensional mapping coordinates of the standard superior crop samples over the years at each sampling node as control anchor points; use the actual crop development stage value as a parameterized time variable, and use the B-spline curve fitting algorithm or polynomial smoothing interpolation algorithm to fit the discrete control anchor points in the low-dimensional feature space into a parameterized standard trajectory curve that is continuously differentiable everywhere. The processing unit employs a spatial Euclidean distance metric algorithm. Using the 32-dimensional co-evolutionary parameter vector acquired at the current moment as a spatial free point, it calculates the Euclidean distance between this vector and each point on the 32-dimensional continuous standard trajectory curve, and solves for the minimum distance. This allows it to calculate the shortest projection distance between the current state and the standard trajectory curve in the feature space. The unit then locates the nearest projection anchor point on the standard curve and extracts the developmental stage value associated with that anchor point, outputting it as the current plant's developmental stage value.

[0072] Cross-correlation analysis was used to quantify the asynchronous delay in the transformation from endogenous metabolism to extrinsic phenotype in plants, and time axis offset compensation was performed accordingly to achieve benchmark alignment of physical and physiological characteristics on the same developmental phase plane. Subsequently, an unsupervised pre-trained nonlinear autoencoder was used to reduce the dimensionality of the aligned high-dimensional data and output it as a co-evolutionary parameter. The shortest spatial Euclidean distance of this parameter to the historical standard trajectory curve was calculated, thereby transforming the discrete growth period of the crop into numerical output of developmental stages.

[0073] Further, the steps for determining the physiological maturity state include: constructing a continuous time-series curve based on the aroma rate of the area; dynamically increasing the filtering threshold along the signal amplitude axis from low to high; performing lower level set topological filtering to extract the time series of the aroma rate of the area; generating a persistent barcode by calculating the connected component evolution characteristics of the sequence at different scales; retrieving the core feature item with the longest lifespan from the persistent barcode; marking the topological extinction time corresponding to the core feature item as the global maximum point of the synthesis rate; and determining the physiological maturity state of the fruit based on the time position of the maximum point.

[0074] First, the collected multi-source data is preprocessed to extract the basic analysis sequence. Discrete data points of regional fragrance rate within the continuous monitoring period are retrieved from the multi-source fusion data center and arranged in chronological order of sampling time to construct a one-dimensional fragrance rate time series signal. To eliminate the interference of short-term high-frequency noise caused by environmental abrupt changes on subsequent topological evolution calculations, a smoothing filter operation is performed on the one-dimensional time series signal using a moving average window to obtain the basic fragrance rate sequence. The time span of the moving average window is set based on the physiological metabolic cycle characteristics of the target crop. In this example, the length of the moving average window is set to the number of sampling points corresponding to 3 to 5 calendar days. Alternatively, based on the Fourier transform spectrum analysis of historical fragrance rate data, the cutoff frequency of high-frequency environmental noise is extracted, and the width of the moving average window is dynamically determined accordingly to ensure that low-frequency trend characteristics are obtained while filtering out short-term high-frequency interference caused by diurnal temperature differences.

[0075] Secondly, a continuously varying traversal threshold is set to construct a lower-level set topological filtering process. The global minimum and global maximum values ​​of the smoothed aroma rate base sequence are obtained. Within this numerical range, multiple continuously increasing filtering threshold levels are divided at equal intervals according to a preset step size. The preset step size is calculated by taking the local standard deviation of the aroma rate base sequence during the unstable period and setting the step size to a value slightly larger than the local standard deviation to avoid false topological connected components due to fluctuations in sensor accuracy. Alternatively, the step size can be set to a value slightly larger than the theoretical detection error extreme value of the data acquisition sensor (such as an infrared spectrometer) to avoid false topological connected components due to fluctuations in the sensor's own accuracy. The computer system scans layer by layer from the lowest threshold to the highest threshold. For each set threshold, all data points in the time series whose aroma rate values ​​are less than or equal to the current threshold are extracted. As the scanning threshold gradually increases, the number of data points that meet the conditions increases continuously on the time axis and becomes adjacent, forming several continuous data segments based on the time axis extension. These continuous data segments are defined as connected components.

[0076] Subsequently, the evolution trajectory of the aforementioned connected components is tracked to generate a persistent barcode. This invention employs a lower level set filtering strategy: during the scanning process where the threshold increases layer by layer from low to high, when the time series passes through a local signal trough, a new independent connected component is formed, and the threshold at this time is recorded as the topology birth scale. As the threshold increases, each connected component expands on the time axis. When two adjacent connected components converge and merge at the signal peak, the threshold at this time is recorded as the extinction scale according to the life cycle priority rule in topology, and the time coordinate corresponding to the convergence point is recorded as the topology extinction moment. During the scanning process where the threshold increases layer by layer, when the time series passes through a local minimum value of the birth rate, a new independent connected component will be formed at that point. The threshold at which this component is generated is recorded as the topology birth scale, and the time corresponding to the trough is recorded as the birth moment. As the threshold continues to increase, each independent connected component expands to both sides on the time axis. When two adjacent connected components converge and contact at a local maximum value on the time axis, a fusion operation is performed. During fusion, the lifecycle priority rule in topology is applied to retain connected components with lower birth thresholds, mark the other connected component as terminated, and force the threshold corresponding to the intersection and fusion to be recorded as the extinction scale. The time coordinate corresponding to the intersection point is recorded as the topological extinction time of the terminated connected component. The birth and extinction scales of all connected components are transformed into interval line segments to generate a persistent barcode set representing the trajectory of the synthesis rate fluctuation.

[0077] Finally, persistent barcodes are selected and time nodes of physiological maturity are marked. The generated persistent barcode set is traversed, and the difference between the extinction scale and the birth scale of each interval segment is calculated. This difference represents the lifespan of the connected component in the threshold space. By sorting the values, the interval segment with the largest lifespan value is retrieved as the core feature. The topological extinction time bound to this core feature is extracted. Since the extinction of a connected component must be mapped to the signal peak under the lower level set filtering framework, and the vanishing point of the core feature with the longest lifespan corresponds to the main peak with the largest fluctuation amplitude in the entire sequence, this topological extinction time can be objectively marked as the global maximum point of the aroma substance synthesis rate. Finally, the time position of this global maximum point is mapped to a specific calendar date to determine the state node where the fruit has reached the target physiological maturity.

[0078] This technical solution uses a dynamic filtering mechanism to remove high-frequency environmental noise while retaining low-frequency trend signals. Combined with an adaptive traversal step size setting rule, it reduces the loss of topological details and the generation of spurious components caused by subjective experience settings or inherent sensor errors. On this basis, it uses the life cycle priority rule in topology to transform discrete time-series signals into persistent barcodes, quantifying the hierarchical evolution relationship of signal extreme points from a global perspective. Finally, it relies on extracting the core feature term with the largest life cycle to lock the synthetic main peak, so that the judgment of crop physiological maturity is no longer directly dependent on the absolute concentration threshold, and has a certain degree of objectivity and stability when facing complex meteorological differences.

[0079] Furthermore, the steps of generating the harvest prescription map and delineating the distribution of harvestable areas include, specifically, as follows: Figure 3 As shown: The acquired developmental stage values ​​and fruit physiological maturity status are mapped to a preset farm geographic information coordinate matrix to obtain a set of spatial nodes with physiological status labels; preset logistics constraints and fruit grading standards are retrieved, and joint screening and priority scoring are performed on the set of spatial nodes to obtain a harvest distribution map; nodes with priority values ​​higher than a preset threshold are extracted from the harvest distribution map, and spatial neighborhood scanning is performed on the nodes to statistically analyze local spatial cluster density, resulting in dense communities representing areas with consistent maturity and concentrated harvesting value; gridded coordinate calibration is performed on the dense communities, and harvesting time labels are assigned according to priority order, outputting the harvest prescription map containing grid spatial coordinates and batch harvesting priorities.

[0080] First, a geographic information coordinate matrix for the farm is constructed, and spatial mapping of physiological states is performed. High-precision positioning equipment is used to acquire the geographic boundary coordinates of the target area, which are then projected and transformed into a unified Cartesian coordinate system. Discretization and grid segmentation are performed on the planar area according to a preset spatial resolution (0.5m × 0.5m actual physical spacing in this example), generating a basic geographic information coordinate matrix for the farm. Subsequently, the developmental stage values ​​and fruit physiological maturity data of each plant or cluster of crops obtained from previous calculations are retrieved. The corresponding GPS spatial coordinates at the time of data collection are extracted, and an inverse distance weighting algorithm is used to bind the aforementioned physiological feature data as multi-dimensional attribute vectors to the nearest grid node in the coordinate matrix. This constructs a set of recognition spatial nodes containing two-dimensional spatial coordinates and physiological state labels. Specifically, when using the inverse distance weighting algorithm, the power exponent parameter of distance decay is set to a constant between 2 and 3 to strengthen the clustering properties of local physiological features; simultaneously, the maximum search radius of the algorithm is set to be equal to or slightly greater than the average row spacing of the crops. If a grid node does not match any actual plant coordinates within its maximum search radius, the grid node is marked as a blank area and will not participate in subsequent priority scoring, thereby avoiding incorrect mapping of uncultivated passageways or wastelands to physiological status labels.

[0081] Secondly, constraints are introduced to perform joint screening and priority scoring on the spatial node set. Pre-set fruit grading standards (the minimum harvestable sugar content or lower fruit diameter limit set in this example) and current logistics constraints (the remaining cold storage capacity or maximum throughput of the sorting line selected in this example) are retrieved from the database. For the spatial node set, threshold filtering is first performed using the fruit grading standards to remove invalid grid nodes that have not reached the physiological maturity threshold. For the retained valid nodes, a weighted scoring function is constructed: the physiological maturity status of the node is extracted, and the reciprocal of the time difference from the end of the optimal harvest period is calculated as the harvest urgency factor; simultaneously, the straight-line distance from the node's coordinates to the nearest logistics distribution point in the farm is extracted, and a spatial decay factor is calculated; the urgency factor and the spatial decay factor are normalized and then weighted and summed to output a quantitative score representing the harvest priority for each valid node, thus generating a preliminary harvest distribution map.

[0082] Finally, a spatial neighborhood scan is performed to extract dense communities and output the final map. A priority filtering threshold is set (in this example, the top 30% of priority scores in the entire orchard are used as a baseline). Candidate high-priority nodes with scores higher than this threshold are selected from the picking distribution map. A density-based spatial clustering algorithm is used, with each candidate high-priority node as the center, and a physical distance representing a reasonable operating radius for manual picking (set to 3 to 5 meters in this example) is defined as the search neighborhood. The number of other high-priority nodes falling into this neighborhood is counted. If this number is greater than the preset minimum threshold for community formation (in this example, 15 consecutive clusters of trees), it is marked as a core cluster area. Core cluster areas with overlapping neighborhoods are further... The boundaries of the clusters are merged to form independent, densely populated regional communities. Isolated high-priority nodes that fail to meet the minimum threshold for community formation and cannot be merged with any core cluster area are identified as scattered early-maturing fruits and removed as spatial noise, excluding them from this centralized harvesting plan. For the boundaries of the irregular densely populated regional communities, the outermost geometric contour surrounding the community is extracted using the convex hull algorithm and mapped back to the basic farm geographic information coordinate matrix. A rasterization scanning transformation mechanism is used to uniformly classify all discrete grids whose center points fall within the contour into the coverage area of ​​the community, thereby achieving grid alignment of the irregular boundaries. Grid alignment is performed on the boundaries of these densely populated regional communities. The average priority score of all nodes within the community is extracted as the global priority of the community. The data, including spatial grid coordinates, cluster area contours, and corresponding global priority highlight rendering layers, is packaged and output as a harvest prescription map that can be directly read by the farm management terminal system.

[0083] In the process of generating prescription maps, this technical solution first isolates cropless areas through a limited spatial interpolation algorithm, ensuring the objectivity of the underlying state label mapping. Second, it establishes a joint screening mechanism with multi-dimensional variable constraints, quantifying and integrating fruit physiological indicators, harvesting time urgency, and logistics transportation distance to output harvesting priorities. Finally, it uses spatial clustering combined with outlier removal logic to focus the operation target on high-value dense communities, and utilizes convex hull algorithm and rasterization scanning mechanism to transform irregular natural boundaries into regular digital grids, outputting a harvest prescription map that can be used for farm management.

[0084] Example 2: To address the asynchronous monitoring issue of "ripe but not fully flavored (sugar / aroma not meeting standards)" in greenhouse-grown "Sunshine Rose" grapes during the secondary swelling and ripening stages, this application introduces an intelligent monitoring method for crop growth cycles based on multi-source data fusion. During the fruit color change and flavor formation stages of Sunshine Rose grapes, the system scans the grape bunches using an indium gallium arsenide image sensor deployed on a mobile inspection platform to obtain infrared reflectivity reflecting the distribution of carbohydrates in the pulp. Simultaneously, a polarimetric synthetic aperture radar mounted on a UAV releases microwave pulses to penetrate dense foliage and obtain the microwave scattering coefficient reflecting the density of the berries. Furthermore, a pulse amplitude modulation fluorometer is used to acquire leaf photosynthetic yield at fixed intervals, combined with continuous time-series accumulated temperature collected by an off-ground meteorological sensing terminal, providing core input for subsequent metabolic modeling.

[0085] To address the compact nature and susceptibility to compressive stress in Shine Muscat grapes, this study first extracts the structural scattering component from the microwave scattering coefficient. Then, it uses a crop spatial topology prior knowledge base to perform mapping calculations, defining the physical morphological boundary set for ear volume expansion. Simultaneously, it extracts spectral absorption features from the infrared reflectance to define the physiological development boundary set for the fruit. Within a unified feature phase space, logical intersection calculations are performed to lock in the effective developmental domain. The spatial dimension and density properties of steady-state nodes within this domain are then analyzed, directly mapping and outputting the ear enlargement and tissue density of the current region.

[0086] To address the energy competition between sugar and aroma synthesis in Shine Muscat grapes during their ripening period, this embodiment constructs a multi-parameter competitive model. First, real-time leaf photosynthetic yield and time-series accumulated temperature are spliced ​​using tensors to construct a basal metabolic potential energy vector representing the plant's biochemical capabilities. The canopy reflectance index is then extracted as a spatial bias parameter to dynamically generate an environmental transformation hyperplane. Subsequently, this potential energy vector is orthogonally decomposed toward the hyperplane, decoupling two independent signals entering the competitive allocation mapping layer: one inputs the parallel components representing primary metabolism into the time-series prediction network and the first fully connected layer, outputting the fruit sugar content curve and turgor pressure variation rate (a warning of fruit cracking risk) via nonlinear mapping; the other inputs the vertically orthogonal components representing secondary stress metabolism into the second fully connected layer, calculating the aroma rate of monoterpenoid characteristic substances via scalar regression, reflecting the fruit's internal ripening status.

[0087] The geometric bounding box in a three-dimensional virtual spatial coordinate system is defined by the fruit bunch enlargement. Tissue density is assigned as a quality weight to discrete sampling anchor points, and a signed distance field is constructed using implicit neural functions. By performing spatial differentiation calculations on this distance field, the Gaussian curvature tensor under the Jacobian matrix is ​​solved, thereby mapping and outputting the regional structural characteristics reflecting the surface tension and geometric developmental qualitative changes of Sunshine Rose fruit. Simultaneously, the sugar content curve, turgor pressure change rate, and aroma rate are aligned along the time axis, and discrete wavelet transform is used to separate the approximate coefficients corresponding to the long-term growth trend and the detailed coefficients corresponding to environmental stress, outputting global memory features.

[0088] By constructing a morphological-physiological asynchronous delay metric, look-ahead offset compensation is performed on the structural features of the area to achieve benchmark alignment of the two-dimensional features on the same developmental phase plane. The aligned features are used to extract co-evolutionary parameters that characterize the intensity of co-development, and distance measurement is performed with the standard trajectory space to output the developmental stage values. By extracting the aroma rate time series, the evolutionary features of its connected components at different scales are calculated to generate a persistent barcode, and the topological extinction moment of the core feature term is marked as the global maximum point of the synthesis rate, thereby determining the optimal physiological maturity state of the fruit.

[0089] The developmental stage values ​​and physiological maturity status are mapped to a preset farm geographic information coordinate matrix. Combined with logistics constraints and grading standards, priority scoring is performed. Spatial nodes with priority values ​​higher than the threshold are automatically extracted. Through spatial neighborhood scanning, dense communities with consistent maturity and concentrated harvesting value are identified. A harvest prescription map containing gridded spatial coordinates and batch harvesting priorities is output.

[0090] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent monitoring of crop growth cycle based on multi-source data fusion, characterized in that, include: The infrared reflectance, microwave scattering coefficient, leaf photosynthetic yield, canopy reflectance index, and time-series accumulated temperature of the plants in the area were obtained. Based on the physical morphological constraints of microwave scattering coefficient and the physiological state constraints of infrared reflectivity, the ear enlargement and tissue density are output through joint boundary intersection addressing. By inputting leaf photosynthetic yield, time-series accumulated temperature, and canopy reflectance index into a multi-parameter competitive model, the sugar content curve, turgor pressure change rate, and aroma rate of the fruit were obtained. Based on the enlargement of the ears and the density of the tissue, a signed distance field is constructed to extract the structural features of the regions; based on the sugar content curve of the plants in the regions, the turgor pressure change rate and the aroma rate of the regions, global memory features are extracted through multi-scale time-frequency domain feature decomposition and multi-level feature fusion. Based on the structural features of the plant area and the global memory features, the developmental stage of the plant is calculated by temporal phase difference. Based on the aroma rate, a continuous temporal curve is constructed, the topological evolution features of the connected components of the aroma rate of the plant area are extracted, the global maximum point of the aroma rate synthesis rate is calibrated, and the physiological maturity status of the fruit is determined. Based on the developmental stage values ​​and the physiological maturity of the fruit, a harvest prescription map is generated to delineate the distribution of harvestable areas.

2. The intelligent monitoring method for crop growth cycle based on multi-source data fusion according to claim 1, characterized in that, The steps for obtaining the infrared reflectance, microwave scattering coefficient, leaf photosynthetic yield, time-series accumulated temperature, and canopy reflectance index of plants in a given area include: using an image sensor to obtain the infrared reflectance of the plants in the given area; using a polarimetric synthetic aperture radar to release microwave pulse signals for canopy penetration detection to obtain the microwave scattering coefficient of the plants in the given area; using an ultra-hyperspectral imager to perform continuous spectral scanning detection of the canopy of the plants in the given area to obtain the leaf photosynthetic yield; using a multispectral camera to take overhead images to obtain the canopy reflectance index; and using an off-ground meteorological sensing terminal to collect air temperature data at a continuous sampling frequency and obtain the time-series accumulated temperature of the plants in the given area through integral calculation based on a reference temperature.

3. The intelligent monitoring method for crop growth cycle based on multi-source data fusion according to claim 1, characterized in that, The process for outputting the ear enlargement and tissue density of plants in a given area includes: extracting the structural scattering features of the microwave scattering coefficient; performing feature space mapping calculations on the structural scattering features using a preset crop spatial topology prior knowledge base to delineate the physical morphological boundary set of plant development; extracting the spectral absorption features of the infrared reflectance; performing feature space mapping calculations on the spectral absorption features using a preset prior knowledge base to delineate the physiological development boundary set of plant development; performing logical intersection calculations on the physical morphological boundary set and the physiological development boundary set in a unified feature phase space to obtain the effective development domain; parsing the steady-state node attributes contained within the effective development domain; and directly mapping and outputting the spatial dimension parameters and density parameters corresponding to the steady-state node attributes as the ear enlargement and tissue density of the plants in the given area.

4. The intelligent monitoring method for crop growth cycle based on multi-source data fusion according to claim 1, characterized in that, The multi-parameter competitive model includes: Metabolic potential encoding layer: Based on a preset historical time sliding window, extract the photosynthetic yield of the leaves and the accumulated temperature of the time series for the consecutive N days before the current time node to construct a one-dimensional time vector, and perform tensor splicing operation to construct the basic metabolic potential vector in the feature space; Hyperplane generation layer: Extract the scalar features of the canopy reflectance index, use the scalar features as spatial bias parameters and scale scaling factors, dynamically adjust the mapping scale of the environmental transformation matrix, generate the hyperplane corresponding to the transformation matrix in the feature space, and define the orthogonal projection boundary of the basal metabolic potential energy through the geometric position of the hyperplane. Orthogonal projection decoupling layer: The basic metabolic potential energy vector is spatially geometrically aligned with the hyperplane, and an orthogonal decomposition calculation oriented towards the hyperplane is performed on the basic metabolic potential energy vector to separate the parallel component vector and the vertical residual vector. Competitive allocation mapping layer: receives parallel component vectors and vertical residual vectors; inputs the parallel component vectors into a preset temporal prediction network and a first fully connected layer, and outputs the sugar content curve and turgor pressure variation rate; inputs the vertical residual vectors into a preset second fully connected layer, and outputs the aroma production rate.

5. The intelligent monitoring method for crop growth cycle based on multi-source data fusion according to claim 1, characterized in that, The steps for extracting the structural features of the region include: receiving the ear enlargement and tissue density of the plants in the region; defining the geometric bounding box boundary in the three-dimensional coordinate system using the ear enlargement, and performing grid sampling on the space within the boundary to generate discrete coordinate points; assigning the tissue density as a quality weight to the discrete coordinate points to construct a set of discrete sampling anchor points; based on the spatial distribution topology of the discrete sampling anchor points, continuously fitting the coordinate space using an implicit neural function to construct a signed distance field; performing differential calculation on the signed distance field to solve for the continuous gradient distribution of the distance field, and calculating the Gaussian curvature tensor based on the Jacobian matrix of the gradient distribution; performing feature dimension concatenation operation on the continuous gradient distribution and the Gaussian curvature tensor, and performing dimensionality reduction encoding to map and output the structural features of the region.

6. The intelligent monitoring method for crop growth cycle based on multi-source data fusion according to claim 1, characterized in that, The steps for obtaining the global memory features include: resampling and aligning the sugar content curves, turgor pressure change rate, and aroma rate of the plants in the region on the same time axis, and performing tensor splicing to obtain a metabolic time series signal; setting a multi-scale time sliding window for different phenological stages of the covered crop, and performing mean and variance operations on the metabolic time series signal to obtain the trend mean and fluctuation variance terms of physiological fluctuations; and using a discrete wavelet transform operator to perform multi-level decomposition and downsampling processing on the metabolic time series signal to separate the approximate coefficients corresponding to the growth trend and the detailed coefficients corresponding to the environmental fluctuations. The following steps are performed: Discrete integral operations are executed on the feature channels corresponding to the sugar content curve and aroma rate in the approximation coefficients to obtain long-term cumulative sugar and aroma substance values; the feature channels corresponding to the turgor pressure variation rate in the detail coefficients are extracted and their sum of squared amplitudes is calculated to obtain short-term turgor pressure fluctuation energy values; the trend mean term, fluctuation variance term, cumulative sugar and aroma substance values, and short-term turgor pressure fluctuation energy values ​​are concatenated into a vector, and then input into a preset principal component analysis model and linear mapping matrix for subspace projection. After removing redundant dimensions, the global memory features are output.

7. The intelligent monitoring method for crop growth cycle based on multi-source data fusion according to claim 1, characterized in that, The steps for outputting the current plant developmental stage value include: inputting the region structural features and the global memory features from the continuous time series into a preset time sliding window; performing partial derivative calculations on the corresponding multidimensional feature vectors and feature latent states to output the morphological change gradient and metabolic evolution rate respectively; constructing a morphological-physiological asynchronous delay metric based on the developmental phase difference between the morphological change gradient and the metabolic evolution rate; using the morphological-physiological asynchronous delay metric to perform look-ahead offset compensation on the time axis of the region structural features to achieve benchmark alignment between physical morphological features and physiological metabolic features on the same developmental phase plane; performing multidimensional topological mapping on the aligned region structural features and the global memory features to extract co-evolutionary parameters; measuring the distance between the co-evolutionary parameters and the preset standard crop variety growth and development trajectory feature space, and outputting the developmental stage value of the plant in the region based on the measurement deviation.

8. The intelligent monitoring method for crop growth cycle based on multi-source data fusion according to claim 1, characterized in that, The steps for determining the physiological maturity state include: constructing a continuous time-series curve based on the aroma rate of a specific area; extracting the time series of the aroma rate of the area by dynamically adjusting the filtering threshold along the signal amplitude axis; generating a persistent barcode by calculating the connected component evolution characteristics of the time series at different scales; retrieving the core feature item with the longest lifespan from the persistent barcode; marking the topological extinction time corresponding to the core feature item as the global maximum point of the synthesis rate; and determining the physiological maturity state of the fruit based on the time position of the maximum point.

9. The intelligent monitoring method for crop growth cycle based on multi-source data fusion according to claim 1, characterized in that, The steps of generating the harvest prescription map and delineating the distribution of harvestable areas include: mapping the acquired developmental stage values ​​and fruit physiological maturity status to a preset farm geographic information coordinate matrix to obtain a set of spatial nodes with physiological status labels; retrieving preset logistics constraints and fruit grading standards, performing joint screening and priority scoring calculation on the set of spatial nodes to obtain a harvest distribution map; extracting nodes with priority values ​​higher than a preset threshold from the harvest distribution map, and performing spatial neighborhood scanning on the nodes to statistically analyze local spatial clustering density to obtain dense communities representing areas with consistent maturity and concentrated harvesting value; performing gridded coordinate calibration on the dense communities, assigning harvest time labels according to priority order, and outputting the harvest prescription map containing grid spatial coordinates and batch harvesting priorities.

Images